| The safe and reliable operation of dynamic processes is not only the basis for guaranteeing industrial production capacity,but is also closely related to the safety and stability of society.In the chemical,metallurgical,power grid,and other equipment and key infrastructures that involve dynamic processes,there are many new demands,nevertheless,the research on basic theories and algorithms is still the cornerstone of innovation and development.With the dynamic processes being increasingly complex,the reliability of safe monitoring systems becomes essential.Any minor theoretical flaws or algorithmic loopholes may lead to serious accidents.In fact,many world-shocking safety incidents and accidents in recent years have exposed safety issues,prompting us to continuously improve the monitoring theory and methods and to optimize solutions for monitoring systems.In the data-driven framework,this thesis discusses the scientific questions regarding both process operational safety and information security.The weaknesses of the existing methods and the root causes of the problems are revealed through theoretical analysis,and on this basis,solutions and design approaches are proposed correspondingly.The relationship between several types of fault diagnosis methods applicable to dynamic systems is studied.To start with,the quantitative relationship between the residual signal generated by the model-based method and the data-driven implementation method is analyzed.On the basis of existing works,links are established between multiple system descriptions,input/output data,as well as the residual generation approaches based on observers/stable kernel representation/parity space.This illustrates the reasonability of cross-category comparison and acts as the basis for problem transformation.Furthermore,through theoretical derivations and analysis,it is pointed out that the reason for unstable performance of the one-dimensional residual generator lies in the irreversible loss of noise information at the design stage.To retain more useful information,a multi-dimensional residual generator design scheme is proposed,which can effectively improve the stability of the fault detection rate.An in-depth study is conducted on the selection of Luenberger equation solutions in the design of multi-dimensional residual generator observer,aiming to answer how many sets of solutions are needed,which solutions should be selected,and the weights of different sets of solutions,so as to avoid blindness in design.According to the correspondence between the solutions of the Luenberger equations and the parity vectors,these problems are transformed to the study in parity space.To find a parity vector so that the corresponding residual signal is least affected by noise,a parity vector parameterization method and an iterative parameter update algorithm are proposed,which minimizes the residual expectation in the fault-free scenario and guarantees the convergence of the variance.The proposed optimization method can avoid severe performance degradation in the cases of tiny faults.Furthermore,to deal with strong nonlinearity in the monitored systems,an approach based on the least square support vector machine is proposed,which expands the applicable range of the optimization method.In order to judge whether the detected faults will affect the key performance indicators(KPIs),research is carried out in the framework of multivariate statistical analysis.The dual principal component analysis-based fault detection approach is reviewed,and the reason why it is not suitable for dynamic processes is analyzed.Then,a novel recursive realization approach is proposed for fault detection.With theoretical analysis,it is proved that the proposed algorithm can significantly reduce the amount of online computation and memory usage.To deal with the influence of non-Gaussian data distribution on the test statistics and to determine a reasonable threshold,the kernel density estimation method is used.The proposed approach can be applied to the KPI-oriented fault detection tasks for dynamic systems,and shows the best overall performance among a series of recursive and non-recursive approaches.Towards the security monitoring problems related to data networked transmission in dynamic processes,the research is focused on protecting the data transmission process from eavesdropping attacks and integrity attacks.From the attackers’ point of view,several popular types of attacks are analyzed,among which false data injection attacks(FDIAs)can stay concealed against detection.A problematic statement in a recently published paper is pointed out.To correct such,the condition of the stealthiness(undetectability)of FDIAs with respect to the system topology is proposed and proved.It is verified that in certain scenarios,attackers can manipulate the systems’ states trajectories by controlling the states of the attacking system without affecting the outputs.From another viewing angle,secure data transmission is studied aiming at designing defending systems.An integrated data-driven design framework against eavesdropping attacks and integrity attacks is put forward.Furthermore,based on the correlation among variables,a modularized encryption/decryption approach and a trustworthiness judgment method are proposed. |
